Objaverse-PBR-render

PBR rendered condition videos for Ink3D — a 3D texture generation pipeline using video generative models.

📄 Paper: Ink3D: Sculpting 3D Assets with Extremely Complex Textures via Video Generative Models
💻 Code: github.com/YueHan99/Ink3D.TextureGen
🤗 Model: Yuehavingfun/orbitpainter-single

This dataset provides pre-rendered geometry condition videos (position, normal, albedo, RGB, depth, mask) for ~23,000 Objaverse models, rendered from horizontal (H) and vertical (V) camera orbits (120 cameras each, 120 frames at 1024×1024).

Directory Structure

h/videos_curve/{bucket}/{uuid}/
├── albedo.mp4       # Base color
├── position.mp4     # World-space position map
├── normal.mp4       # Normal map
├── rgb.mp4          # Shaded color (white background)
├── depth.mp4        # Depth map
└── mask.mp4         # Alpha mask

v/videos_curve/{bucket}/{uuid}/
├── albedo.mp4
├── position.mp4    
├── normal.mp4
├── rgb.mp4
├── depth.mp4
└── mask.mp4

• Buckets: 121 directories (000-000 to 000-120)
• Samples: ~23,617 models per track
• Total: ~140K video files, ~142 GB

Render Your Own Data

To render condition videos from your own GLB models, use the Ink3D Render module:

Reconstruction

Verify dataset quality by baking rendered videos back to textured PBR GLB using the Ink3D reconstruction pipeline:

conda activate trellis2
python bake_pbr.py --sha256 "000-000/{uuid}" --video_path "./albedo_h.mp4" --output_dir ./output

Produces bake_1024_a.glb with baseColor + metallicRoughness textures from the rendered videos and camera parameters.

Source

Rendered from Objaverse 1.0 GLB models using Blender 4.0+ with CYCLES GPU.

Citation

@article{han2026ink3d,
  title   = {Ink3D: Sculpting 3D Assets with Extremely Complex Textures via Video Generative Models},
  author  = {Han, Yue and Li, Chong and Liu, Zhening and Huang, Cong and Deng, Fang and Liu, Yong and Wei, Fangyun and Lu, Yan},
  journal = {arXiv preprint arXiv:2607.01222},
  year    = {2026},
}